- No due dates exist—everything is "yesterday"
- No artificial blockers; only physical/fundamental constraints matter
- Attempt 1-year projects in 1 month; you'll finish in 2 months (still 6x faster)
- Every time estimate is based on assumptions—audit and eliminate them for 2x+ speedups
- Only 3 management layers: ICs → co-founders/managers → Elon
- Fuzzy team boundaries: anyone can fix anything, merge immediately
- "Live by the sword, die by the sword" - ask for responsibility, deliver or leave
- Binary decision responses: "no, that's dumb" or "why isn't it done already?"
- No formal onboarding or team assignments—just laptop + badge
- Challenge "perceived limitations"—most are false, especially around speed/latency
- 2-8x improvement is achievable on anything built in the last 10 years by removing stack overhead
- Daily (sometimes multiple times daily) model iterations from pre-train
- Delete-first methodology: remove complexity, add back only what's proven necessary
- One person + 20 AI agents can rebuild core production APIs
- Digital equivalent of Optimus robot: automate any keyboard/mouse/screen work
- Target: 8x faster than human (not larger models—faster ones)
- No software adoption required—directly emulates human computer interaction
- Counter to industry: small fast models, not big slow reasoning models
- Scaling from 1,000 to 1 million human emulators is "not very big" challenge
- 4M+ Tesla vehicles in North America with capable computers
- 70-80% of time sitting idle with networking, cooling, power
- Pay owners to lease compute time for human emulators
- More capital efficient than AWS/Oracle/Nvidia hardware
- Zero buildout required—purely software implementation
- Hardware is their biggest competitive advantage
- Colossus data center built in 122 days
- "Carnival permit" loophole used for fastest permitting
- 80+ mobile generators for seamless grid failover
- Power scales by megawatts in milliseconds; multi-layer capacitor/battery/generator system
- Elon calls Nvidia, gets patches next day—weeks of back-and-forth compressed to hours
- "Engineers, just engineers"—problem solvers from any background
- Look for simplicity: 10-line solution > 200-line AI-generated solution
- Include incorrect requirements in interview—expect candidates to push back
- Talent density is extreme; first place Sully had to work hard to keep up
- 3-person iOS team serving millions of users
- Everyone is an engineer, including sales team
- Goes "fire to fire," unfucks whatever problem exists
- Feedback is bimodal: very high level (strategy) or very low level (latency, compute)
- Proof required, not opinion—run experiments
- Updates timelines daily based on new information
- Proactively asks "how can I help?" at end of meetings
- Calibrated his timeline estimates through experience—now much more accurate
- Start from revenue/outcome target, work backwards to physical requirements
- Identify future bottlenecks years in advance
- Physical requirements determined last, not first
- Focus on single core metric that drives financial/physical return
- Everyone sits in same building, reachable by walking to desk
- No docs—things move too fast; trying to auto-generate with Grok
- Managers who code (though less now with 100+ reports)
- Smart people who are "very nice and helpful," not stuck-up
- War room operations for 4+ months; outgrew it, moved to the gym
- Sleeping pods and bunk beds for surge nights
- $2.5 million value per commit to main repo
- 5 commits = $12.5 million value added in one day
- iOS team of 3 people for massive user base
- <8 non-engineers at early xAI
- Smaller models iterate faster AND perform faster—compounding advantage
- Information degrades through management layers ("compression")—minimize layers
- AI happily writes 200 lines; humans must find the 10-line solution
- One brain can do more now that AI handles the typing
- Virtual employees already being tested internally (people don't know they're AI)
- Generalization surprising better than expected—untrained tasks done flawlessly